fix(ir): the wheel says which passage it is and nothing about what is in it, and Spanish comes off the air

[joaocarvoli]

A team turned the passage wheel and heard fourteen lines per language that we wrote from each map's pericope-title, a summary rather than a name, so the wheel told the team what it was about to find before it had chosen anything; P07 handed over the redeemer. The file says of itself that nothing in it has been approved to be spoken to a team and that no validator sees it. A Spanish tablet, meanwhile, met every failure of the room in lines from a second draft file, and opened sessions in a language Marcia has not approved. Her ruling of 4 September settles both: nothing about a passage may be said outside the map, and Spanish is not offered until she decides to offer it.

The wheel now says which passage it is and nothing about what is in it. Every line is the map's own H1 reference, the way her menu takes it: # P07 — Ruth 2:17-23 becomes Ruth 2:17-23, with the book named in the language being spoken, Rute in Portuguese from the label catalogue. The verse dash is copied from the canon rather than normalized (an en dash through P05, a hyphen from P06 on), because a line that disagrees with the map it came from is the thing this stops. The lock is a test that reads the fourteen maps and the catalogue and demands that every line in the file, in every language, be its map's reference with the book renamed, so a rewrite that carries story again has nothing to agree with; the expected lines are transcribed from the maps, not derived from the file under test. The old no-digit assertion is inverted: a reference is numbers, and what must stay out is the canon's own file id.

Spanish comes out of what a team can reach. ROOM_LANGUAGES loses es, and since normalize is the one gate the wire passes through, the wheel, the session open and the voice route answer 400 for it, beside the French they already refused. LANGUAGE_NAMES and the Spanish voice keep their entries so a session already written as es is read back rather than turning into a 500. The wheel's Spanish block goes. The fail-safe supplements are no longer globbed from the directory: they are named off ROOM_LANGUAGES, so the Spanish draft stays exactly where it is, banner intact, unreachable from any path that speaks, and returning es to the tuple is the whole of turning it back on. The passage-lines DRAFT banner stays where it is.

Scope 6 asked for the 2026-08-10 approval of the Portuguese supplement as an artifact beside the file. There was nothing to archive: the header's "Reviewed and approved by Marcia (2026-08-10)" was written thirteen seconds after one line of chat relaying that she had agreed, the Portuguese it vouches for had been written fifty-eight seconds earlier in the same session, no commit records an approval, her repository has none on that date, and on 4 September she asked for the text rather than confirming from memory. So the header now says what it has, an oral approval relayed by João with her written confirmation pending, and the record sits beside the file in _fail_safe_pt_supplement_provenance.md. The lines themselves are untouched; whether the Portuguese B, C and E stand is her call once she has the text. João: confirm your recollection of 10 August in the morning.

Verified against this worktree, host-run backend, speech through a local stub: the wheel in Portuguese and in English speaks six lines (P07 still gated), each the bare reference, Rute 1:1–5 … Rute 2:8-16 and Ruth 1:1–5 … Ruth 2:8-16; es and fr answer 400. The app half is the next PR, stacked on this one.

---

Verificação

critério	teste	resultado	falsificado
C1 · cada linha da roda é a referência do H1 do seu mapa, com o nome do livro na língua da sessão	tests/test_internalization_room_passages.py (28 literais transcritos dos catorze mapas)	passa	sim
C2 · nenhuma linha do arquivo, em nenhuma língua, carrega mais que a referência (lê os mapas e o catálogo)	mesmo arquivo	passa	sim — linha reescrita com história
C3 · es não é reivindicado: normalize devolve nada; roda, sessão e voz respondem 400 (es, es-419, fr)	tests/test_internalization_room_language.py	passa	sim
C4 · nenhuma linha nem asset es alcançável de passages.py / fail_safe.py; suplementos nomeados por ROOM_LANGUAGES	tests/test_internalization_room_fixed_lines.py	passa	sim
C5 · sessão persistida em es é lida de volta sem 500	mesmo seam	passa	sim
C6 · proveniência do suplemento pt escrita ao lado do arquivo	inspeção (não é comportamento)	—	—

Gates: uv run ruff check ., ruff format --check ., mypy app/, boots, alembic heads (uma), python scripts/check_doctrine.py verdes; suíte inteira 2602 passed, 4 skipped, 1 xfailed. Revisor interno (joao-review): 0 achados sobreviventes. 6 commits, 10 arquivos, +212 −90.

Muda o que a equipe ouve na roda (referência falada em vez de frase): não entra sem o teste de UI do João (portão de teste manual, de manhã). ENG-787 (P07 re-vendor) stacks on the same line rule.

Ticket: ENG-759.

🤖 Generated with Claude Code

https://claude.ai/code/session_01AucpnrSDAiaKpcsBCNH7C2

[linear-code]

ENG-759

[little-henok]

Line 22 says "Nothing here has been approved to be spoken to a team", and passages.py:82 speaks these lines to a team on the wheel. This PR treats that exact sentence as disqualifying for a file in the same directory: test_internalization_room_fixed_lines.py:104 asserts it against the Spanish draft before asserting the draft reaches no mouth. Pick one — drop the banner because a bare map reference needs no approval, narrow it to what is actually pending, or keep the file off the wheel until she rules. I cannot read her 4 September ruling from here; whether the reference form is approved needs Henok.

[joaocarvoli]

Narrowed it. What is settled is that the lines are the maps' own references, which is her 4 September rule applied and the form her menu shows; what still awaits her word is whether the wheel speaks a reference aloud at all. The banner now says that, and the ticket's instruction not to delete it quietly is kept. de2b1fc.

[little-henok]

en is FLOOR (languages.py:23), and first falls back through utterances to the untagged authored English (fail_safe.py:57), so this case can no longer separate "honoured the session language" from "ignored it and took the floor" — with es it could. Both assertions hold either way. The fallback is one dropped argument away: language_code: str = FLOOR, back_translation.py:992. Not firing today, the pt case at line 402 still catches it. The helper above states the standard itself: "no case could fail for the reason it exists". Run this one with language="pt" and swap the assertions, or drop it and say the pair needs a third language.

[joaocarvoli]

Right: with English being the floor, that case could not fail for the reason it exists. Dropped it; the Portuguese case stands alone, and the docstring of the commit says a second contrast needs a third claimed language. de2b1fc.

[little-henok]

Read the whole diff: the language gate, the fail-safe loader, the rewritten passage lines and the tests. Two findings, and neither breaks the room today — one banner that contradicts what the wheel does with the file it sits on, one test case that stopped being able to fail when es became en. Not checking the provenance record's account of 10 August against the session logs, and not the pronunciation of the new references either; that one is your morning UI pass. Those two are the whole list.

[little-henok]

Second pass, ledger from pass 1 re-read against de2b1fc. Both points are closed. The four banner lines are the only lines added since c86194ea, and they cannot reach a mouth: passage_lines.py:10-11 matches only ### Pnn-xx headings and - "…" bullets, and the banner sits above the first heading; the STATUS: DRAFT string asserted at tests/test_internalization_room_fixed_lines.py:104 belongs to _fail_safe_es_supplement.md, not this file. Nothing new raised and nothing left open from me. Not covering the app half stacked behind this. Approval is Henok's.